Release v0.7.0

Vector Inspector 0.7.0 — April 5, 2026

A major milestone release introducing full text and image ingestion, multimodal embeddings, and inline file previews.
You can now build real multimodal collections, inspect their embeddings, and run text→image semantic search directly inside Vector Inspector.

---

🚀 Highlights

✓ Multimodal Ingestion (Images + Documents)

Import images, text files, PDFs, Word docs, and source files into any collection.
Images are embedded with CLIP (512‑dim).
Documents are chunked and embedded with MiniLM (384‑dim).

✓ Text → Image Semantic Search

You can now type a natural‑language query and retrieve matching images from your collection using CLIP’s shared embedding space.

✓ Inline File Previews

Images and text files now show thumbnails or text snippets directly in the details pane and item dialog, making it easy to verify ingestion and debug metadata.

✓ Robust, Production‑Ready Ingestion Pipeline

Chunking, duplicate detection, partial‑ingest recovery, and detailed logging ensure ingestion is deterministic, resumable, and transparent.

---

🧩 Ingestion

• Image ingestion pipeline using CLIP (openai/clip-vit-base-patch32, 512‑dim)
• Document ingestion pipeline using sentence-transformers (all-MiniLM-L6-v2, 384‑dim)
• Import via “Import Images…” and “Import Documents…” in the Tools menu
• Paragraph‑aware chunking for documents (1000 chars default) with chunk_index, chunk_total, parent_id, and file metadata
• Three‑way duplicate detection:
  • new files ingested
  • fully-present files skipped
  • partially-ingested files automatically cleaned and re‑ingested
• Re-ingest file… context menu option for single-file overwrite
• Lazy loading of heavy dependencies with clear install guidance
• Ingestion dialog shows filename + progress (e.g. “3 of 42”)
• Collections auto-refresh after ingestion
• Per-file log entries restored (Ingested image: …, Ingested document: …)
• Telemetry: ingestion.started and ingestion.completed with full metrics
• New collections created at ingestion time via CollectionService
• Backends without configurable vector size show read-only dimension label

---

🖼️ File Preview

• New File Preview section in the inline details pane
• Image thumbnails:
  • 160×120 inline
  • 320×240 in item dialog
• Text previews:
  • 30 lines / 2 KB inline
  • 100 lines / 8 KB in dialog
• Right-click actions: Open and Reveal in Explorer/Finder/Files
• Double-click image → open in OS viewer
• Metadata table now shows a 📎 icon for rows with previewable files
• Preview detection via find_preview_paths() with safe fallbacks
• Text detection via mimetypes.guess_type + null-byte sniff
• Collapsed state persisted in settings

---

🛠️ Bug Fixes & Stability

• Fixed _flush() in ingestion pipelines to correctly detect and raise on failed writes
• Fixed CLIP crash on tiny images (<3×3 px) with a clear error message
• Fixed embedding nesting: _l2_normalize now flattens to 1D
• Truncated long error strings to avoid log flooding
• Silenced noisy third‑party loggers (chromadb, sentence_transformers, etc.)
• Fixed UI crash when metadata contained non‑JSON‑serializable types (e.g. uuid.UUID) via a new metadata sanitizer

---

🎉 Summary

0.7.0 transforms Vector Inspector into a true multimodal semantic debugging tool.
You can now ingest real documents and images, inspect their embeddings, preview their contents, and run text→image semantic search — all with a stable, production‑grade ingestion pipeline.

---

Installation

Download the .whl or .tar.gz file from the Assets section below, then install with:

pip install <your-filename.whl>
# or
pip install <your-filename.tar.gz>

After installation, run the application with:

vector-inspector

For more details and troubleshooting, see the README.

What's Changed

• 0.7.0 - Image and text file Ingestion, Document preview by @anthonypdawson in #29

Full Changelog: v0.6.2...v0.7.0